Deep Learning-Based Classification of Diabetic Retinopathy Severity Using DenseNet201 and SVM with Cubic Kernel

Diabetic retinopathy (DR) is a significant contributor to global blindness and is a progressively deteriorating eye condition that is a consequence of diabetes mellitus. It is imperative to accurately classify the severity of diabetic retinopathy (DR) in order to promptly diagnose and treat the condition. In this investigation, we employed DenseNet201, an effective deep learning framework, to extract deep features from retinal images. The extracted features were subsequently fed into a Support Vector Machine (SVM) classifier with a cubic kernel to perform multi-class classification of Diabetic Retinopathy (DR) into five stages: mild, moderate, proliferative, severe, and non-diabetic. The model achieved favorable results during validation, with an Area Under the Curve (AUC) that ranged from 0.99 to 1 and an accuracy rate of 99%. However, the performance experienced a slight decline when evaluated on a distinct test dataset, resulting in an accuracy of 82% and AUC values comprising from 0.94 to 0.99. In real-world implementations of DR classification systems, the challenges of attaining generalizability are emphasized by the discrepancy between the validation and test outcomes.